# Lincense: Academic Free License (AFL) v3.0
#
import sys
from time import sleep
import numpy as np
from prosper.utils.parallel import pprint
from prosper.utils.datalog import dlog, StoreToH5, TextPrinter
# Parameters
rf_shape = (26, 26)
H = 16
# Configure Data-Logger
dlog.set_handler(('T', 'S', 'C'), TextPrinter)
# And GO!
D = rf_shape[0] * rf_shape[1]
Wshape = (H,D)
i = 0
for T in np.linspace(0., 20, 50):
i = i + 1
pprint( "%i th iteration..." % i)
W = np.random.normal(size=Wshape)
dlog.append_all( {
'T': T,
'W': W,
'S': np.sin(T),
D = ts_size**2 # dimensionality of observed data
# Approximation parameters for Expectation Truncation (It has to be Hprime>=gamma)
Hprime = 5
gamma = 3
# Import and instantiate a model
discriminative = False
if discriminative:
model = DBSC_ET(D, H, Hprime, gamma)
else:
model = BSC_ET(D, H, Hprime, gamma)
# Configure DataLogger
print_list = ('T', 'L', 'pi', 'sigma')
dlog.set_handler(print_list, TextPrinter) # prints things to terminal
txt_list = ('T', 'L', 'pi', 'sigma')
dlog.set_handler(txt_list, StoreToTxt, output_path +
'/results.txt') # stores things in a txt file
h5_list = ('T', 'L', 'pi', 'sigma', 'W')
dlog.set_handler(h5_list, StoreToH5, output_path +
'/results.h5') # stores things in an h5 file
# Choose annealing schedule
from prosper.em.annealing import LinearAnnealing
anneal = LinearAnnealing(20) # decrease
anneal['T'] = [(0, 5.), (.8, 1.)]
anneal['Ncut_factor'] = [(0, 0.), (0.5, 0.), (0.6, 1.)]
# anneal['Ncut_factor'] = [(0,0.),(0.7,1.)]
# anneal['Ncut_factor'] = [(0,0.),(0.7,1.)]
anneal['W_noise'] = [(0, np.std(ts) / 2.), (0.7, 0.)]
# Disgnostic output
pprint("=" * 40)
pprint(" Running bars experiment (%d parallel processes)" % comm.size)
pprint(" size of training set: %d" % N)
pprint(" size of bars images: %d x %d" % (size, size))
pprint(" number of hiddens: %d" % H)
pprint(" saving results to: %s" % output_path)
pprint()
# Generate bars data
my_data = model.generate_data(params_gt, N // comm.size)
# Configure DataLogger
print_list = ('T', 'Q', 'pi', 'sigma', 'N', 'MAE')
store_list = ('*')
dlog.set_handler(print_list, TextPrinter)
dlog.set_handler(print_list, StoreToTxt, output_path + '/terminal.txt')
dlog.set_handler(store_list, StoreToH5, output_path + '/result.h5')
model_params = model.standard_init(my_data)
if 'anneal' in params:
anneal = params.get('anneal')
else:
# Choose annealing schedule
anneal = LinearAnnealing(50)
anneal['T'] = [(0, 2.), (.7, 1.)]
anneal['Ncut_factor'] = [(0, 0.), (2. / 3, 1.)]
anneal['anneal_prior'] = False
# Create and start EM annealing
pprint(" Running %d parallel processes" % comm.size)
pprint("=" * 70)
H = 2 * D2 # number of latent units
D = D2**2 # total size of image in pixels
my_N = N // comm.size
# Some sanity checks
assert Hprime <= H
assert gamma <= Hprime
assert D == D2**2
# Configure DataLogger
print_list = ('T', 'pi', 'sigma')
dlog.set_handler(print_list, TextPrinter)
# Invent some ground truth parameter models
params_gt = {'W': 10 * generate_bars_dict(H), 'pi': 2. / H, 'sigma': 1.0}
# Use model to generate data
model = BSC_ET(D, H, Hprime, gamma)
my_data = model.generate_data(params_gt, my_N)
model_params = model.standard_init(my_data)
# Choose annealing schedule
anneal = LinearAnnealing(50)
anneal['T'] = [(15, 1.), (-10, 1.)]
anneal['Ncut_factor'] = [(0, 0.), (2. / 3, 1.)]
anneal['anneal_prior'] = False